targets_type autoencoder_type batch_size
exp_id
3 10_Targets Over_dim 64
16 10_Targets normal_dim 32
20 Mnist Over_dim_iteration 32
33 10_Targets normal_dim_iteration 32
60 Mnist Over_dim_tied_iteration 64
65 10_Targets Over_dim_tied_iteration 32
68 Noisy Over_dim 64
  • Batch size doesn't change the manifold significantly
  • Training iteration (train the network with predictions) does influence the manifold, mostly rotation and the cluster formation it becomes slightly worse through the iterations. With Noisy data iteration (train the network with predictions) appears to change nothing.
  • Tied weights also seem to make the clustering worse.
  • Over_dim_tied_iteration shows in contrast to Over_dim for 10_Targets worsening. This is a combination of the worsening in Over_dim_tied and Over_dim_iteration.
  • For Noisy data Over_dim and normal_dim don't present a big difference. Although normal dim produces worse reconstructions
  • For Mnist data Over_dim_iteration and normal_dim_iteration don't present a big difference. Although normal dim produces worse reconstructions